Record: Long Context + All Optimizations submission#166
Open
chinesepowered wants to merge 1 commit intoopenai:mainfrom
Open
Record: Long Context + All Optimizations submission#166chinesepowered wants to merge 1 commit intoopenai:mainfrom
chinesepowered wants to merge 1 commit intoopenai:mainfrom
Conversation
Combines best training techniques (seq_len=2048, Muon momentum 0.99, conservative LRs, longer warmdown) with all SOTA eval/quantization tricks (sliding window, FP16 embed, Overtone init, Muon WD, 10 layers). https://claude.ai/code/session_01D1CQCz3TCExUmWTVDivu3R
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Submission:
records/track_10min_16mb/2026-03-20_LongCtx_SlidingWindow_FP16Emb_10L_AllOpts/Strategy: Merge the two strongest approaches that haven't been combined yet:
Key changes from current SOTA:
train_seq_len: 1024 → 2048 (2x more context per training token)train_batch_tokens: 524K → 393K (more optimizer steps per wallclock)matrix_lr/scalar_lr: 0.04 → 0.02 (reduces quantization gap)muon_momentum: 0.95 → 0.99 (stronger gradient smoothing)warmdown_iters: 2500 → 3600 (smoother weights for quantization)tied_embed_lr: 0.10 → 0.06 (more stable embedding training)The Seq4096 submission showed ~0.02 BPB improvement from training alone (without any eval tricks). The SOTA's sliding window eval adds ~0.03 BPB improvement. Combined, this could target ~1.155-1.165 BPB, pending validation on 8xH100 hardware.